TransEdge: a lightweight transformer-based anomaly detection framework for edge devices

The rapid growth of the Internet of Things (IoT) and the increasing interconnectivity of industrial systems are driving a critical need for anomaly detection to be performed locally on edge devices. However, a significant gap exists between the computational power required by advanced deep learning models and the limited resources of edge hardware, often forcing a compromise between detection capability and on-device feasibility. To bridge this gap, we first introduce the EM-AT and its variants (EM-AT-bin), which enhance the Anomaly Transformer (AT) by integrating the Expectation–Maximization (EM) algorithm to enable fully automated, data-driven threshold determination. The EM-AT model with a Bayesian information criterion (BIC) estimator achieves the highest detection performance across four public datasets (i.e., SWaT, WADI, HDFS, and OpenStack), with $$F_{1}$$ -scores of 96.32%, 92.47%, 98.90%, and 99.61%, respectively. Building on the EM-AT and its variants, we present Q-EM-AT (Quantized-EM-AT) and its variants (Q-EM-AT-bin), edge-optimized and quantized variants that leverage mixed-precision quantization to substantially reduce computational and memory overhead while preserving detection accuracy. Finally, we propose TransEdge, a lightweight edge anomaly detection framework that adopts Q-EM-AT as its core detector to balance detection performance and computational resource consumption. Comprehensive experiments show that TransEdge significantly reduces resource consumption while maintaining competitive detection performance, achieving $$F_{1}$$ -scores of 96.18% on SWaT, 92.36% on WADI, 98.65% on HDFS, and 99.43% on OpenStack.